Closed pangsu0613 closed 5 years ago
The usage has no much difference with the original mscnn. We have provided the pretrained model for cascade-mscnn. Just use https://github.com/zhaoweicai/mscnn/blob/master/examples/kitti_car/run_cascademscnn.m for running the inference.
Thank you for your quick response! Have you done any experiment on KITTI pedestrian and cyclist benchmark?
The goal of the cascade r-cnn is for high quality object detection, but KITTI pedestrian and cyclist use AP50 for evaluation. So I haven't tested on them yet.
Ok, I see, thanks for your explanation, but I don't understand the AP50 part, do you mean IoU=0.5?. The Cascade-RCNN includes a sequence of detectors with increasing IoU. But since KITTI uses IoU=0.5 for evaluation, which is a pretty low metric discussed in the paper, so that's the main reason, am I correct ?
Yes, you are correct. When you look at the paper, you can find Cascade R-CNN doesn't have much improvement when the evaluation metric is loose, e.g. AP50 (IoU=0.5).
Ok, I think I get it. Thank you so much for your help.
Hello! Thank you for sharing your newest work! Could you please provide some instructions on how to use the new cascade RCNN for KITTI 2d object detection benchmark? Many thanks!!